1. Field of the Invention
The present invention generally relates to computer systems, and more particularly to a method of efficiently scheduling snoop operations between caches in a multiprocessor computer system.
2. Description of Related Art
The basic structure of a conventional multiprocessor computer system 10 is shown in FIG. 1. Computer system 10 has several processing units, two of which 12a and 12b are depicted, which are connected to various peripheral devices, including input/output (I/O) devices 14 (such as a display monitor, keyboard, graphical pointer (mouse), and a permanent storage device or hard disk), memory device 16 (such as random access memory or RAM) that is used by the processing units to carry out program instructions, and firmware 18 whose primary purpose is to seek out and load an operating system from one of the peripherals (usually the permanent memory device) whenever the computer is first turned on. Processing units 12a and 12b communicate with the peripheral devices by various means, including a generalized interconnect or bus 20, or direct memory access channels (not shown). Computer system 10 may have many additional components which are not shown, such as serial, parallel, and universal system bus (USB) ports for connection to, e.g., modems, printers or scanners. There are other components that might be used in conjunction with those shown in the block diagram of FIG. 1; for example, a display adapter might be used to control a video display monitor, a memory controller can be used to access memory 16, etc. The computer can also have more than two processing units.
In a symmetric multi-processor (SMP) computer, all of the processing units are generally identical, that is, they all use a common set or subset of instructions and protocols to operate, and generally have the same architecture. A typical processing unit includes a processor core 22 having a plurality of registers and execution units, which carry out program instructions in order to operate the computer. An exemplary processing unit includes the PowerPC(trademark) processor marketed by International Business Machines Corp. The processing unit can also have one or more caches, such as an instruction cache 24 and a data cache 26, which are implemented using high speed memory devices. Caches are commonly used to temporarily store values that might be repeatedly accessed by a processor, in order to speed up processing by avoiding the additional latency of loading the values from memory 16. These caches are referred to as xe2x80x9con-boardxe2x80x9d when they are integrally packaged with the processor core on a single integrated chip 28. Each cache is associated with a cache controller (not shown) that manages the transfer of data and instructions between the processor core and the cache memory.
A processing unit can include additional caches, such as cache 30, which is referred to as a level 2 (L2) cache since it supports the on-board (level 1) caches 24 and 26. In other words, cache 30 acts as an intermediary between memory 16 and the on-board caches, and can store a much larger amount of information (instructions and data) than the on-board caches can, but at a longer access penalty. For example, cache 30 may be a chip having a storage capacity of 512 kilobytes, while the processor may be an IBM PowerPC(trademark) 604-series processor having on-board caches with 64 kilobytes of total storage. Cache 30 is connected to bus 20, and all loading of information from memory 16 into processor core 22 must come through cache 30. Although FIG. 1 depicts only a two-level cache hierarchy, multi-level cache hierarchies can be provided where there are many levels (L3, L4, etc.) of serially connected caches.
In a multi-level cache, if a copy of a value is in every level of the cache, the cache hierarchy is referred to as being xe2x80x9cinclusive.xe2x80x9d It is not necessary, however, to keep a copy of each value in the lower levels, and an inclusivity bit field may be added to the caches to indicate whether or not the cache is inclusive. For example, a three-level cache structure might provide an L3 cache which was not inclusive, such that a value residing in the L2 cache might not be present in the L3 cache. In this example, if an L2 cache issues a read command for a value that is not present in any of the caches of that processing unit, it can be passed to that L2 cache without (necessarily) loading it into the L3 cache.
In an SMP computer, it is important to provide a coherent memory system, that is, to cause write operations to each individual memory location to be serialized in some order for all processors. By way of example, assume a location in memory is modified by a sequence of write operations to take on the values: 1, 2, 3, 4. In a cache coherent system, all processors will observe the writes to a given location to take place in the order shown. However, it is possible for a processing element to miss a write to the memory location. A given processing element reading the memory location could see the sequence 1, 3, 4, missing the update to the value 2. A system that implements these properties is said to be xe2x80x9ccoherentxe2x80x9d. Nearly all coherency protocols operate only to the granularity of the size of a cache block. That is to say, the coherency protocol controls the movement of and write permissions for operand data or instructions on a cache block basis, and not separately for each individual memory location.
There are a number of protocols and techniques for achieving cache coherence that are known to those skilled in the art. All of these mechanisms for maintaining coherency that the protocols allow only one processor to have a xe2x80x9cpermissionxe2x80x9d that allows a write operation to a given memory location (cache block) at any given point in time. As a consequence of this requirement, whenever a processing element attempts to write to a memory location, it must first inform all other processing elements of its desire to write the location and receive permission from all other processing elements to carry out the write.
To implement cache coherency in a system, the processors communicate over a common generalized interconnect (i.e., bus 20). The processors pass messages over the interconnect indicating their desire to read from or write to memory locations. When an operation is placed on the interconnect, all of the other processors xe2x80x9csnoopxe2x80x9d (monitor) this operation and decide if the state of their caches can allow the requested operation to proceed and, if so, under what conditions. There are several bus transactions that require snooping and follow-up action to honor the bus transactions and maintain memory coherency. The snooping operation is triggered by the receipt of a qualified snoop request, generated by the assertion of certain bus signals. Instruction processing is interrupted only when a snoop hit occurs and the snoop state machine determines that an additional cache snoop is required to resolve the coherency of the offended sector.
This communication is necessary because, in systems with caches, the most recent valid copy of a given block of memory may have moved from the system memory 16 to one or more of the caches in the system (as mentioned above). If a processor (say 12a) attempts to access a memory location not present within its cache hierarchy, the correct version of the block, which contains the actual (current) value for the memory location, may either be in the system memory 16 or in one of more of the caches in another processing unit, e.g. processing unit 12b. If the correct version is in one or more of the other caches in the system, it is necessary to obtain the correct value from the cache(s) in the system instead of system memory.
For example, consider a processor, say 12a, attempting to read a location in memory. It first polls its own L1 cache (24 or 26). If the block is not present in the L1 cache, the request is forwarded to the L2 cache (30). If the block is not present in the L2 cache, the request is forwarded on to lower cache levels, e.g., the L3 cache. If the block is not present in the lower level caches, the request is then presented on the generalized interconnect (20) to be serviced. Once an operation has been placed on the generalized interconnect, all other processing units snoop the operation and determine if the block is present in their caches. If a given processing unit has the block requested by processing unit in its L1 cache, and the value in that block is modified, and any lower level caches also have copies of the block, then their copies are stale, since the copy in the processor""s cache is modified. Therefore, when the lowest level cache (e.g., L3) of the processing unit snoops the read operation, it will determine that the block requested is present and modified in a higher level cache. When this occurs with an in-line cache structure, the L3 cache places a message on the generalized interconnect informing the processing unit that it must xe2x80x9cretryxe2x80x9d it""s operation again at a later time, because the actual value of the memory location is in the L1 cache at the top of the memory hierarchy and must be retrieved to make it available to service the read request of the initiating processing unit.
Once the request from an initiating processing unit has been retried, the L3 cache begins a process to retrieve the modified value from the L1 cache and make it available at the L3 cache, main memory or both, depending on the exact details of the implementation. To retrieve the block from the higher level caches, the L3 cache sends messages through the inter-cache connections to the higher level caches, requesting that the block be retrieved. These messages propagate up the processing unit hierarchy until they reach the L1 cache and cause the block to be moved down the hierarchy to the lowest level (L3 or main memory) to be able to service the request from the initiating processing unit.
The initiating processing unit eventually re-presents the read request on the generalized interconnect. At this point, however, the modified value has been retrieved from the L1 cache of a processing unit and placed into system memory, and the read request from the initiating processor will be satisfied. The scenario just described is commonly referred to as a xe2x80x9csnoop pushxe2x80x9d. A read request is snooped on the generalized interconnect which causes the processing unit to xe2x80x9cpushxe2x80x9d the block to the bottom of the hierarchy to satisfy the read request made by the initiating processing unit.
Thus, when a processor wishes to read or write a block, it must communicate that desire with the other processing units in the system in order to maintain cache coherence. To achieve this, the cache coherence protocol associates with each block in each level of the cache hierarchy, a status indicator indicating the current xe2x80x9cstatexe2x80x9d of the block. The state information is used to allow certain optimizations in the coherency protocol that reduce message traffic on the generalized interconnect and the inter-cache connections. As one example of this mechanism, when a processing unit executes a read it receives a message indicating whether or not the read must be retried (i.e., reissued later). If the read operation is not retried, the message usually also includes information allowing the processing unit to determine if any other processing unit also has a still active copy of the block (this is accomplished by having the other lowest level caches give a xe2x80x9csharedxe2x80x9d or xe2x80x9cnot sharedxe2x80x9d indication for any read they do not retry). Therefore, a processing unit can determine whether any other processor in the system has a copy of the block. If no other processing unit has an active copy of the block, the reading processing unit marks the state of the block as xe2x80x9cexclusivexe2x80x9d. If a block is marked exclusive it is permissible to allow the processing unit to later write the block without first communicating with other processing units in the system because no other processing unit has a copy of the block. Therefore, it is possible for a processor to read or write a location without first communicating this intention onto the interconnection, but only where the coherency protocol has ensured that no other processor has an interest in the block.
The foregoing cache coherency technique is implemented in a specific prior art protocol referred to as xe2x80x9cMESI.xe2x80x9d In this protocol, a cache block can be in one of four states, xe2x80x9cMxe2x80x9d (Modified), xe2x80x9cExe2x80x9d (Exclusive), xe2x80x9cSxe2x80x9d (Shared) or xe2x80x9cIxe2x80x9d (Invalid). Under the MESI protocol, each cache entry (e.g., a 32-byte sector) has two additional bits which indicate the state of the entry, out of the four possible states. Depending upon the initial state of the entry and the type of access sought by the requesting processor, the state may be changed, and a particular state is set for the entry in the requesting processor""s cache. For example, when a sector is in the Modified state, the addressed sector is valid only in the cache having the modified sector, and the modified value has not been written back to system memory. When a sector is Exclusive, it is present only in the noted sector, and is consistent with system memory. If a sector is Shared, it is valid in that cache and in at least one other cache, all of the shared sectors being consistent with system memory. Finally, when a sector is Invalid, it indicates that the addressed sector is not resident in the cache.
While these various cache operations ensure proper coherency and allow for the sharing of memory between the different processing units, there are still several problems regarding the manner in which these operations are carried out. For example, certain system bus commands (i.e., kill-type commands such as the PowerPC(trademark) DClaim or read-with-intent-to-modify commands) must be snooped by all caches in order to complete. If any snooper on the system bus retries the requested command (e.g., due to a directory busy condition) the master must re-issue the request. This condition can create poor performance (or livelock situations) because the snoopers are not synchronized, and there is no assurance that snoopers with limited bandwidth will able to snoop each operation.
Some systems may provide a second system bus to allow two addresses to be transmitted. In the situation wherein all L3 caches can snoop only one address (of the two address that may arrive on the two system buses), and both of these system buses are saturated, the L3 must uses an algorithm to randomly pick the snoop from one of the buses. This approach leads to very poor odds for getting a kill-type operation through cleanly (i.e. without being retried). Assuming that, when two addresses are presented on the system buses at the same time to a given L3 cache, then there is only a 50% chance that the snooper will choose the kill operation. For larger SMP systems (e.g., 64-way), the snooping traffic is exponentially complicated, and the odds that all 64 L3 snoopers will choose the address for the pending kill operation (0.5)64, i.e., extremely unlikely.
Another example involves the refresh time where dynamic random-access memory (DRAM) circuits are used for an L4 cache. If each L4 directory refresh time occurs every 512 clock cycles and lasts four cycles, and each L4 cache is performing the refresh asynchronously with respect to other L4 caches, then the odds of an address being sent while the L4 cache is refreshing are (4/512)64, again extremely unlikely. This factor will thus lead to further retries and delays.
Typically, the L3 and L4 directories are single ported such that in any given cycle, either a read or write (but not both) can be performed. In many cases, these write operations take multiple cycles such that once initiated, a directory read (i.e. snoop) cannot be immediately performed. This situation is similar to that described above in that, if many snoopers are performing directory writes and these writes happen asynchronously to each other, snoops have a high probability of getting retried by a directory that has already committed to a write and is in a busy condition.
When a directory, like the L4 DRAM directory described above, is off-chip, it is often partitioned into several banks (e.g., bank0,1,2,3). Bank selection is typically achieved through an address hash. For example, if the cache line size is 128 bytes and the address is a 64-bit addr(0:63), then the bank selection may use address bits 55:56 to determine which bank to use. Due to the nature of a DRAM bank, a new read may only be issued every four system bus cycles. Since none of the L4 directories are synchronized or aligned, a master device that issues an address which hashes into a certain bank may find that for each attempt, it runs into an L4 snooper that is in a bank busy condition(i.e. must retry) because it has already committed to a snoop that occurred to that same bank one or two cycles earlier.
In light of the foregoing, it would be desirable to devise an improved method of scheduling snoop traffic in a multiprocessor computer system. It would be further advantageous if the method could utilize existing infrastructure without adding excessive complexity or cost, yet still ensure that certain instructions were snooped within a limited time period.
It is therefore one object of the present invention to provide an improved method of handling cache operations in a multiprocessor computer system.
It is another object of the present invention to provide such a method that efficiently schedules snoop operations between processing units and their caches.
It is yet another object of the present invention to provide a computer system which ensures that certain operations (e.g., kill-type operations) are executed within a limited time frame and without repeated retry attempts.
The foregoing objects are achieved in a method of communicating between caches of different processing units in a multiprocessor computer system, generally comprising the steps of synchronizing snoop operations of the caches, and then issuing a cache operation during a cycle which is selected based on the manner in which the caches have been synchronized. Each master device (cache controller) is aware of when these synchronized, or aligned, snoop tenures occur, so the masters can target these cycles for certain types of requests that are sensitive to snooper retries, such as kill-type operations. In one implementation, wherein there are multiple system buses for increased bandwidth, the synchronization sets up a matching priority scheme, e.g., where all snoopers choose a first bus over the second bus in even-numbered cycles, and choose the second bus over the first bus in odd-numbered cycles. Alternatively (or additionally), the DRAM memory of the snooper""s directory can be synchronized relative to the DRAM refresh window, such that all snooper directories are always performing the refreshes at the same time. In another aspect of the present invention, the system may be initialized by a system bus command which instructs all L3 snoopers to start a counter, to create appropriate windows of time during which a directory will not receive write operations (i.e., the directory is reserved for only read-type operations). In this manner, masters can again issue critical operations when it is known that the snooper""s directories are aligned to take snoops. During this xe2x80x9cgoldenxe2x80x9d cycle, all snoopers postpone any non-snoop activity, such as write-type operations. The invention may also be implemented in a cache hierarchy which provides memory arranged in banks. In such a system, a system bus initialization command may be used to trigger the snoopers into starting a rotating counter as to when they will accept snoops to a certain bank. The invention is not limited to any particular type of instruction as a candidate for special treatment using the aligned or golden cycles, and the functionality may be hardware or software programmable.
The above as well as additional objectives, features, and advantages of the present invention will become apparent in the following detailed written description.